Transfer learning technology-based facial expression recognition method

Abstract

The invention discloses a transfer learning technology-based facial expression recognition method. The method includes the following steps that: real-time facial data collection is carried out, facialimages are recognized and intercepted, grayscale processing is performed on the facial images, LBP features are extracted from faces, average faces are compared with each other, so that domains to which each of test samples belong are determined, whether a corresponding transfer model exists in a model library is checked, if the transfer model exists in the model library, an image file is inputted into the model for prediction, otherwise, other steps are further executed; and a transfer learning method for source domain sample sampling is adopted to perform model training and prediction, under a condition samples of a target domain are insufficient, the small number of the samples of the target domain are adopted to guide the source domain sample sampling, if relevant data samples are correctly judged, new samples are marked, and source domain data retained by sampling and target domain data together participate in the next round of the training of a supervised machine learning model,and the newly trained model can be used to predict facial expression categories in the target domain in the future. The method has high recognition accuracy in cross-domain facial expression classification in real environments.

Description

technical field [0001] The invention belongs to the technical field of facial data processing, in particular to a facial expression recognition method based on transfer learning technology. Background technique [0002] Migration learning is a very active research direction in the current machine learning field. It is proposed to solve some machine learning problems where the probability distribution of data in the source domain and the target domain are different, or the source domain and the target domain are related but the tasks are different. However, there are also some shortcomings in transfer learning, most of which are widely studied in academia, and there are relatively few applications of transfer learning in the industry. [0003] Facial expression recognition is an important research field in artificial intelligence and pattern recognition disciplines. Its purpose is to recognize facial expression categories in real-world scenes through feature engineering and ...

AI Technical Summary

Problems solved by technology

[0007] There are two fundamental problems in the existing transfer learning: on the one hand, the amount of data in the target domain is small, and it is necessary to find auxiliary domain data for instance migration, while the data in the domain that can assist the target domain data is relatively single, which is likely to cause data distribution and target data. The distribution of domain data is not the same, resulting in negative migration

On the other hand, if you want to continue to develop the target domain data, marking the unlabeled data in the target domain requires a lot of labor and expert knowledge, and consumes a lot of resources, and it is impossible to judge whether the labeled data is correct or not.

[0008] (2) The source domain training samples have limitations

[0009] Most of the existing facial expression data model training is based on some benchmark datasets. These datasets are unique and are often photographed or processed under specific conditions. The famous facial expression benchmark datasets such as the KDEF dataset, whose shooting background It is single and the tone of the shooting is basically uniform, but the background in the real scene is often not single and has a variety of tones, so the model trained with a single data set often produces deviations in predicting expressions in specific environments

[0010] (3) The data model is fixed in the software product, and it is impossible to update and iterate the model according to the specific production environment

[0011] At present, commonly used software products based on machine learning or pattern recognition often combine training data training to obtain a better model in the early stage of production, but the training data of the model may have data distribution inconsistencies with the data in the actual environment. If the model is updated and iterated, then the software products based on the model may gradually not adapt to the new environment

For some software industries, repeated training of new models for new environments greatly consumes manpower and material resources, and the solidified data model is not conducive to cost savings for enterprises

[0012] (4) Migration learning technology is rarely used in practical application products

Images