Meta-learning-based zero-sample and small-sample identification photo anomaly detection method

Abstract

The invention discloses a meta-learning-based zero-sample and small-sample identification photo anomaly detection method. The method comprises the following steps: constructing a fine-grained mode training task and a test task; constructing an anomaly detection network based on meta-learning; performing meta-training on the network structure by adopting a fine-grained mode training task to obtainan anomaly detection model; and finally, outputting a normalized detection result by the detection model according to the input task sample, and setting a threshold value to compare and judge whetherthe identification photo to be detected is qualified or not. According to the method, a self-adaptive step updating and feedback optimization strategy is embedded into the meta-learner, so that the meta-learner accurately locates abnormal attributes of samples, and the detection performance of the model under the condition of zero / small samples is improved; whether the photos uploaded by the handler meet the identification photo standard or not can be quickly and effectively judged, the cost and time for manual screening and labeling are saved, and the detection rate is increased.

Description

technical field [0001] The invention belongs to the field of computer vision and deep learning, and more specifically relates to a zero-sample and small-sample ID photo anomaly detection method based on meta-learning. Background technique [0002] At present, with the construction and popularization of the Internet of Things and smart cities, automation and intelligence have been integrated into all aspects of people's lives. At the same time, with the simplification of functional departments, self-handling of various certificates has been applied to urban life scenarios. In order to quickly and effectively apply for a certificate, in addition to correctly filling in personal information, it is also necessary to upload a photo that meets the requirements. Therefore, it is very important to test the uploaded ID photos to determine whether the ID photos are qualified. [0003] At present, there are mainly two detection methods for ID photos: (1) manual detection; (2) detecti...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the problems that the detection method based on deep learning does not have a large number of detailed labeled samples, the samples are scarce, and the cost of manual labeling samples is high in actual application scenarios, a meta-learning-based zero-sample and small-sample ID photo anomaly detection method is proposed, which can be used in Automatically switch task forms for learning under two tasks of zero sample and small sample, which can quickly and effectively judge whether the photos uploaded by the handler meet the certificate standard, saving the cost and time of manual screening

At the same time, it solves the problem of poor detector training effect due to insufficient training samples that are fully labeled in actual application scenarios, improves the detection rate and saves the cost of manual labeling

Images