Image data analysis method based on two-dimensional-incremental random weight network

Abstract

The invention discloses an image data analysis method based on a two-dimensional-incremental random weight network, and the method comprises the steps: obtaining an image sample set, and setting an initial parameter of a two-dimensional-incremental random weight network model; according to the characteristic that the network residual needs to be gradually reduced in the model construction process, a supervision mechanism is established, and a candidate hidden layer node pool is generated; in the candidate hidden layer node pool, adding the hidden layer node corresponding to the fastest model residual drop into the current model; and after the new hidden layer nodes are added, a global optimization algorithm is adopted to obtain the output weight of the whole network. And when the node number of the network reaches the preset hidden layer node number or the network residual error meets the expected precision, ending the construction process of the whole network, and finally obtaining a two-dimensional-incremental random weight-based network model which can be used for optimizing the image processing effect. The method not only has good learning and generalization performance, but also can show good application potential in two-dimensional image data analysis.

Description

technical field [0001] The invention relates to the technical field of image classification, in particular to an image data analysis method based on a two-dimensional-incremental random weight network. Background technique [0002] With the rising wave of deep learning, neural networks with mature learning algorithms and strong representation capabilities are widely used in data analysis. A fully connected neural network usually takes a vector input as input to the input layer. In image data modeling, although this kind of vector operation can make the dot product calculation between the input weight and the input feasible, it will inevitably lead to two problems: 1) when the training samples are limited, the disaster of dimensionality will occur; 2) The spatial information of the original multi-dimensional input will be destroyed, resulting in unsatisfactory modeling performance. Therefore, it is necessary to build fully-connected neural networks that can directly process...

AI Technical Summary

Problems solved by technology

In image data modeling, although this kind of vector operation can make the dot product calculation between the input weight and the input feasible, it will inevitably lead to two problems: 1) when the training samples are limited, the disaster of dimensionality will occur; 2) The spatial information of the original multi-dimensional input will be destroyed, resulting in unsatisfactory modeling performance

However, the restriction of the supervision mechanism of this method is limited, and it is easy to generate redundant hidden layer nodes, resulting in an incompact structure and increasing the pressure on computer storage.

Images