Coarse-grained parameter regularization method for convolutional neural network

Abstract

The invention discloses a coarse-grained parameter regularization method for a convolutional neural network,. The method includes: stretching each convolution kernel on the same convolution layer in the convolutional neural network into a one-dimensional column vector, and reshaping the column vectors into a two-dimensional weight matrix; calculating the mean value and variance of each column of the weight matrix, and calculating the covariance of any two columns of the weight matrix; calculating a correlation coefficient between any two columns of the weight matrix according to the covariancebetween the two columns of the weight matrix to obtain a correlation coefficient matrix of the weight matrix, and taking the correlation coefficient as a representation of a difference degree betweenconvolution kernels; and calculating a norm of the correlation coefficient matrix and adding the norm into a loss function of the original convolutional neural network as a coarse-grained regular term. According to the method, more expressive features can be extracted by using fewer convolution kernel parameters, so that higher recognition accuracy is obtained.

Description

technical field [0001] The invention relates to a coarse-grained parameter regularization method oriented to a convolutional neural network, and belongs to the cross technical fields of machine learning, deep learning, image processing and the like. Background technique [0002] With the release of large datasets such as ImageNet and COCO, the capabilities of Convolutional Neural Networks (CNN) have been brought into full play. Although CNN has made cross-age progress in important tasks such as image classification, target detection, speech recognition, and semantic segmentation, it has also begun to emerge and show great potential in practical applications such as pedestrian detection and license plate recognition. However, due to large The data age and industry Industry 4.0 have not yet begun. People’s data collection in real life, especially in the industrial field, is not enough, and the data collected by different units is also different, which directly limits the appli...

AI Technical Summary

Problems solved by technology

[0004] The technical problem to be solved by the present invention is: the existing L1, L2 regularization performs indiscriminate fine-grained operations on all weight parameters while ignoring the problem of the integrity of the convolution kernel

Images