An Image Recognition Method Based on Improved Focal Loss Function

Abstract

This patent discloses an image recognition method based on an improved Focal loss function. On the basis of the existing Focal loss function, the modulation factor of the Focal loss function is improved, so that the function pays more attention to difficult samples, and the The attention of simple samples is relatively reduced; then, based on the convolutional neural network model based on the Focal loss function, the remaining negative sample sets are predicted, all difficult samples are screened, and divided into N equal parts, which are added to the original training set respectively, N new training sets are formed, then multiple models are trained, and the final predicted image label result is determined according to the voting selection of the N models. On the basis of the original Focal loss function, the present invention pays more attention to difficult samples and improves the generalization ability of the model; in the case of too many negative samples, the original model is used to screen out difficult samples and add them to the training set, not only The accuracy of the model is improved, the available samples are fully utilized, and the amount of computation when training the model is reduced.

Description

technical field [0001] The invention belongs to the technical field of multi-label image recognition based on deep learning, and in particular relates to an image recognition method based on an improved Focal loss function. Background technique [0002] Convolutional Neural Network (CNN) is a neural network specially designed for image recognition problems. CNN can effectively obtain the features of the original image, and it rarely goes through the preprocessing process. Compared with traditional machine learning, the convolutional neural network can automatically extract the image features, which originates from two important characteristics of the convolutional neural network: sparse connection and weights are shared. Sparse connection means that the nodes of the convolutional layer are only connected to some nodes of the previous layer, and are only used to learn local features. Weight sharing means that in the process of scanning the image with the convolution kernel,...

AI Technical Summary

Problems solved by technology

For some methods to solve the sample imbalance, some do not make full use of all samples, resulting in the lack of part of the sample data, and some use all the samples, which makes the calculation amount very large and increases the cost

Images