Method and device for fine tuning of CNN model

Abstract

The embodiment of the invention discloses a method and a device for fine tuning of a CNN (Convolutional Neural Network) model. The method comprises the steps of: performing extraction of image features of all the input sample images based on a CNN model through pre-training, and obtaining various types of image features; for each type of image features, employing a clustering algorithm to performclustering of the image features in the type, and obtaining a cluster center of each cluster; employing a preset target function to calculate errors, wherein the target function is configured to gather the image features in each cluster to the cluster center; and performing reverse direction transmission of the errors in the CNN model, and performing updating of parameters of the CNN model. Therefore, images of dirty samples and images of samples with good quality can be respectively and individually clustered so as to reduce mutual influence therebetween, facilitate convergence of the imagesof the samples with good quality, reduce the influence of the dirty samples on the CNN model and simplify the implementation.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a method and device for fine-tuning a CNN model. Background technique [0002] Convolutional Neural Network (CNN) is one of the mainstream network models in the field of deep learning. A typical CNN model consists of convolutional layers and pooling layers alternately, such as figure 1 As shown, the input layer 101 inputs an image, the convolution layer 102 performs image feature extraction on each local area of ​​the image of the input layer, and the pooling layer 103 samples the image features of the convolution layer to reduce the dimensionality, and then several layers The fully connected layer 104 connects the image features, and the output value of the last hidden layer 105 is the final extracted feature, and the extracted image features are classified by the Softmax classification layer 106, and the prediction score of the category is output. During training, the p...

AI Technical Summary

Problems solved by technology

Some schemes gather the image features of the same category to the feature mean of the category, so that the image features of the same category gather to the f

Images