Image classification method based on direction gradient histogram in combination with pseudo-reverse learning training stack self-encoder

Abstract

The invention discloses an image classification method based on the direction gradient histogram in combination with a pseudo-reverse learning training stack self-encoder. The method comprises steps that (1), the direction gradient histogram (HOG) is utilized to extract image gradient characteristics, the image directional diagram is calculated, and an HOG operator is utilized to count directionalcharacteristics of some overlapped local regions to acquire HOG characteristics of images. Parameters of different HOG operators are set to acquire several HOG characteristics, and these characteristics are fused into high-dimensional characteristic vectors; (2), the pseudo-inverse learning algorithm is utilized to train a stack self-encoder (PILAE), and the fused high-dimensional characteristicsof the previous step are put into the PILAE to continue learning characteristics; and (3) the characteristics learned in the PILAE are put into a classifier for classification. The two-dimensional information of the images can be extracted by the HOG. The pseudo-reverse learning algorithm is a non-iterative method for training multi-layer feedforward neural networks. The method is advantaged in that the proposed model has the better training time than other models, most hyperparameters are determined by the input data and the network structure, and manual setting is not necessary.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and relates to a method for image feature extraction combined with rapid training of a deep neural network, in particular to a model of a direction gradient histogram combined with pseudo-inverse learning and training stack autoencoders for image classification tasks. Background technique [0002] In recent years, the development of deep learning has once again caused an upsurge in artificial intelligence research. Deep learning originated from the artificial neural network model, an artificial intelligence technology in the 1940s. The first neural network model, the linear perceptron, can be trained to classify a certain input vector model. However, some people later pointed out that the function of the simple linear perceptron is limited, and it cannot handle the problem of linear inseparability. The artificial neural network research began to be at a low ebb. Expect. In recen...

AI Technical Summary

Problems solved by technology

[0004] In the Internet era of data explosion, a large amount of unlabeled data is generated every day, and training a deep model often requires a large amount of labeled data. Creating a complete labeled dataset requires a lot of manpower and time.

Images