Image set classification system and method based on manifold deep learning and an extreme learning machine

Abstract

The invention relates to an image set classification system and method based on manifold deep learning and an extreme learning machine. The image set classification system is characterized by comprising a manifold layer, a conversion layer, an orthogonal layer, a projection layer, a pooling layer, an ELM layer and an output layer. The method comprises the following steps of: representing a multi-view subset of the same target object in an input image as a point in a Grassmann manifold by using a manifold layer; secondly, the conversion layer converts an orthogonal matrix in the Grassmann manifold into a low-dimensional matrix through linear mapping; the third orthogonal layer enables the low-dimensional matrix to form a Grassmann manifold, the fourth orthogonal layer enables the Grassmannmanifold to be mapped to Euclidean space through the projection layer, then data of different training branches are fused through the pooling layer, meanwhile, complexity of data feature mapping is reduced, overfitting of training is controlled, and finally training results are output through ELM layer training. The network structure is relatively simple and more effective, the precision is more accurate, and real-time on-line testing can be realized at a learning speed and a testing speed.

Description

technical field [0001] The present invention relates to an image set classification system and method based on manifold deep learning and extreme learning machine. Background technique [0002] In recent years, with the development of mobile Internet technology, the advent of the era of big data has been promoted. The generation of massive data and the effective analysis and mining of these data have become urgent problems to be solved. The deep learning technology represented by CNN has achieved rapid development in target detection and recognition under the premise of large-scale sample data. These algorithms mainly use a single image as the basic analysis unit. In the actual data source, there are a large number of video image sequences or multi-view image sets originating from the same target object, and the images in the image set can reflect the target object from different aspects. [0003] Traditional single-view images usually use Euclidean distance to measure th...

AI Technical Summary

Problems solved by technology

[0003] Traditional single-view images usually use Euclidean distance to measure the similarity between images, without considering the interrelationship of image set objects

Images