Semi-supervised action recognition training method based on frame level and feature level enhancement

The invention discloses a semi-supervised action recognition training method based on frame-level and feature-level enhancement, which comprises two parts of progressive enhancement and multi-head pseudo-tag, so that a model can still learn robust action characterization in a complex and changeable video scene. The progressive enhancement module firstly performs weak/strong enhancement on the video clip at the frame level, and then performs feature disturbance operation at the feature level, so that diversity transformation on the unlabeled video in a wider disturbance space is realized, and the defect that the enhancement transformation effect of a single frame level is limited is overcome. The multi-head pseudo-tag module enables frame-level strong enhancement, frame-level-feature-level weak enhancement and frame-level-feature-level strong enhancement feature and frame-level weak enhancement feature to be aligned, so that the distance between various features is minimized, and representation consistency con