Face recognition method based on cosine loss function

Abstract

The invention discloses a face recognition method based on a cosine loss function. The face recognition method comprises the following steps: preprocessing face image data needing face recognition training to form a training sample set; constructing a CNN network, inputting the training sample set into the CNN network to carry out face recognition convolutional neural network training, and outputting a first face feature code after the training is completed; inputting the training sample set into a Thomson Problem model for solving, and calculating a second face feature code; performing recent matching calculation on the first face feature code and the second face feature code, correcting CNN model parameters according to a recent matching result, and outputting a face recognition result until the recent matching result meets a set threshold value. In the face recognition process, the coding vector of the face feature is adopted to replace the original FC, so that the training of the neural network is not influenced by the size of the training data set, and the face recognition efficiency is improved.

Description

technical field [0001] The invention relates to the technical field of face recognition, in particular to a face recognition method based on a cosine loss function. Background technique [0002] With the blessing of big data, more and more data are available for face recognition training, ranging from tens of thousands to hundreds of thousands to millions. The increase in data has brought us new problems. Large data sets are difficult to train on the current mainstream Loss functions (such as the Softmax series), because this type of loss function requires the last fully connected layer during training. (FC) The number of neurons should correspond to the number of people in the data set. When the number of people trained is hundreds of thousands, the parameter quantity of the last layer is often about 10 times higher than the previous parameter quantity, and the last layer may need to Occupying nearly 1G of video memory, the input BatchSize of such a 12G video card can only...

AI Technical Summary

Problems solved by technology

Although this technical solution overcomes the differences between training and testing by combining SoftMax loss, cosine similarity loss, and Center Loss as the objective function, this method is not suitable for face recognition training with a large amount of sample data. not tall

[0004] For example, the patent application with the application number CN202010188585.X discloses a face recognition method based on cosine loss with unconstrained conditions, which includes the following steps, S1, acquiring an image to be recognized, performing multi-scale transformation on the image to be recognized, and obtaining an image pyramid; S2 1. The image pyramid obtained in step S1 is input to the MTCNN network, and after the MTCNN network processes the image, facial feature points are obtained; S3, according to the facial feature points of step S2, face correction is performed; S4, using the processed data of step S3 Training the Inception-ResnetV1 convolutional neural network, although this technical solution uses the cosine loss function as a supervisory signal to train the classifier model, this method cannot be applied to face recognition training with a large amount of sample data, and the training efficiency is not high

Images