Identification model training method and identification method for video target state

Abstract

The invention discloses a video target state recognition model training method and a recognition method. The model training method comprises the following steps: inputting an image frame into a feature extraction module to obtain high-level features; performing up-sampling on the high-level features through a spatial reasoning module, performing comparison training on the high-level features and significance labels, calculating a first loss function, and enabling the first loss function to be converged to a preset degree through training; and sequentially inputting the high-level features into a ConvLSTM network, a full connection layer and a Softmax layer in a time reasoning module, then obtaining a prediction state of the current target, carrying out comparison training on the prediction state and a state label, calculating a second loss function, enabling the second loss function to be converged to a preset degree through training, and obtaining a video target state recognition model. The target state in the video image can be identified by using the identification model. Through the method of combining time reasoning and space reasoning, the state recognition precision is high, and automatic feature event detection of the video image sequence is realized.

Description

technical field [0001] The invention belongs to the technical field of video image data processing, and more specifically, relates to a recognition model training method and a recognition method of a video target state. Background technique [0002] The image sequence is a record of the actions and states of the target in a continuous period of time. It has coherence in time and space information, and the target has different spatio-temporal characteristics in different states. Taking advantage of this characteristic of images, artificial intelligence technology is used to build image deep learning models to detect and identify the key frame states of image sequences. On the one hand, it is beneficial to the automatic processing and analysis of image data; In a quasi-real-time application, in order to realize fast and automatic target state grasping and evaluation. [0003] However, the state recognition of video key frame is mainly to recognize the action state of the vide...

AI Technical Summary

Problems solved by technology

However, when identifying the key frame state of an object, some joint point information cannot be used as an aid, and the target in some video images only occupies a small proportion of the image. For the key frame of an image sequence with a small target and a large background area The accuracy of state detection is not high

Images