Transfer learning method based on dynamic multi-batch training and color gamut conversion

Abstract

The invention discloses a transfer learning method based on dynamic multi-batch training and color gamut conversion. The method comprises the following steps: 1) constructing a local data set: manually constructing a web crawler to collect a training data set required by a web picture construction experiment; 2) performing color gamut enhancement and pixel-level processing; 3) cleaning data; 4) selecting a model; 5) performing model training: performing fine adjustment on the convolutional neural network structure through feature fusion of the extracted feature description operator and the local data set, performing training by using the data set after color gamut enhancement, and meanwhile, adding dynamic batch processing training to obtain a new usable model. The PCA-based color gamut enhancement algorithm and dynamic multi-batch training provided by the invention have the advantages that the accuracy rate is improved by about 4 percent compared with the current mainstream model without using the technology, and a good direction is provided for the majority of scholars using deep learning research.

Description

technical field [0001] The invention relates to the field of image classification, in particular to a PCA-based color gamut enhancement algorithm for deep learning data sets and a dynamic batch training algorithm for transfer learning. Background technique [0002] In 2014, the VGGnet model invented by the Visual Geometry team at Oxford University won the championship with a 25.3% error rate in the ISLVRC positioning competition; in the same year, Google's Christian Szegedy and others proposed the GoogleNet model and won the 2014 ISLVRC championship. In this way, we can directly learn from the excellent models of the predecessors and apply them to solve new problems, build models faster, train models or achieve better results. In other words, the purpose of transfer learning is to extract knowledge and experience from one or more source tasks, and then apply it to a target field. At the same time, in response to the lack of deep learning model training data sets, relevant s...

AI Technical Summary

Problems solved by technology

Traditional data enhancement is based on the pixel-level expansion of computer graphics. The picture is based on the RGB mode. It is difficult for the computer to intuitively capture the light and shade, hue and outline of the color. Therefore, the improvement of the accuracy of the model recognition by this method is extremely limited. maintain around 2%

Images