Multi-output regression deep network construction method, structure, equipment and storage medium

Abstract

The invention relates to a method for constructing a multi-output regression deep network, comprising the following steps: constructing a first layer of convolutional layer, outputting an input image of any size as an image of a fixed size; constructing a second layer of double convolutional layer, using meta-filtering The filter group forms a double filter bank, and then uses the double filter bank to perform convolutional neural network operations; constructs the third layer and the fourth layer of double convolutional layers, and repeats the operation of the second layer respectively; constructs the fifth layer input layer and the second layer The six-layer output layer stretches the matrix output by the fourth double convolutional layer into a vector as the input of the fifth layer, and uses the full connection of the activation function to perform nonlinear mapping on it; linearizes the multi-output results of the nonlinear mapping return. By constructing a six-layer network structure, the present invention solves the problem of fixed input image size of the traditional convolutional neural network; utilizes the high reliability of DCNN to improve the performance of traditional CNN; adopts the cosine activation function to overcome the problem caused by the random selection of parameters by the kernel approximation method. of chance.

Description

technical field [0001] The invention relates to the field of machine learning and data mining, in particular to the field of computer vision. Background technique [0002] With the explosion of deep learning, neural networks have once again become a hot spot of attention. In the current popular network structures, such as VGG, Google model, etc., the traditional CNN layer is also used, and in the selection of activation functions, ReLU and its deformation, Sigmod function, Tanh function, etc. are its alternatives, among which ReLU is the most The preferred activation function for the model. In the DCNN article proposed by NIPS in 2016, only the number of layers, the number of nodes, and the activation function in the VGG model are applied, except that the convolutional layer CNN is replaced by DCNN. [0003] The traditional multi-output regression model needs to describe the relationship between input and output, as well as the correlation between multiple outputs. At pre...

AI Technical Summary

Problems solved by technology

[0005] The present invention proposes a new neural network structure and construction method for the end-to-end multi-output regression depth network, which solves the problem of fixed input image size of the traditional convolutional neural network by constructing a six-layer network structure; and utilizes the high Reliability, improve the performance of traditional CNN

Images