Convolutional neural network feature extraction method based on principal component analysis

Abstract

The invention discloses a convolutional neural network feature extraction method based on principal component analysis. The convolutional neural network feature extraction method based on principal component analysis includes the steps: selecting the convolutional neural network which has been trained on an Imagenet data set, taking the convolutional neural network as the feature extractor of an image, extracting the feature mapping graph from the output of each pooling layer of the convolutional neural network, taking all the extracted feature mapping graphs of each layer as the deep featuresof the image, utilizing principal component analysis to perform dimensionality reduction on the deep features, and utilizing bilinear interpolation to reset the final result feature mapping graph tothe original image size to obtain the efficient image deep features. The deep features obtained through convolutional neural network feature extraction method based on principal component analysis have rich semantic information of the image, are low in the feature dimension and small in data size, and can be used for various tasks of identification and classification of images.

Description

technical field [0001] The invention belongs to the technical field of image signal processing, in particular to a convolutional neural network feature extraction method based on principal component analysis. Background technique [0002] In image recognition, classification, and detection tasks, traditional machine learning methods are based on low-level manual features of images, that is, features that can be directly extracted from images. Such as color features, texture features, and histogram features. These features can have good results for images with simple content. However, when encountering images with complex content, manual features cannot effectively represent the nature of the image. The model trained with these manual features Less robust. Traditional image features are more for specific problems, and are specially set manually. This method has weak generalization ability, poor portability, and relatively poor accuracy. [0003] With the rise of deep learni...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to provide a convolutional neural network feature extraction method based on principal component analysis, which solves the extraction problem of the original feature vector and makes the feature easier to use in image processing or computer vision tasks

Images