Convolution neural network training method, gesture recognition method, device and apparatus

Abstract

The invention discloses a training method of a convolution neural network. The method includes: firstly, obtaining a gesture image to be trained; according to Mask R-CNN target detection, segmenting and extracting the gesture image to obtain the coordinates of key points corresponding to each gesture in the gesture image; for each key point, performing corresponding identification according to thevisibility of the key point, so as to obtain the marked characteristic information, wherein, the characteristic information comprises the coordinates of the key point and the corresponding visibilitymark; for each gesture image, reducing the dimensionality of the identified feature information based on a manifold learning algorithm, and obtaining the reduced dimensionality feature point distribution image. For each feature point distribution image, according to the combination of corresponding feature points in the feature point distribution image, obtaining the gesture instruction label after the gesture semantic labeling. According to the feature point distribution image and the corresponding gesture instruction label, the initial convolution neural network is trained to obtain the trained convolution neural network, which simplifies the processing complexity and improves the processing efficiency.

Description

technical field [0001] The invention relates to the technical field of information processing, in particular to a convolutional neural network training method, gesture recognition method and device. Background technique [0002] At present, human-computer interaction technology has gradually shifted from computer-centered to user-centered, and it is an interactive technology with multiple channels and multiple media. Gesture is a natural, intuitive and easy-to-learn means of human-computer interaction. From the traditional mouse and keyboard input to the current infrared, wireless, etc., to the human hand directly as the input device of the computer, the communication between man and machine will no longer need the intermediate media, and the user can simply define an appropriate gesture to Control surrounding machines. This makes human-computer interaction more convenient and rich. At present, the more active human-computer interaction mainly includes speech recognition,...

AI Technical Summary

Problems solved by technology

[0004] When the inventor implemented the embodiment of the present invention, the inventor found that in the prior art, gestures are various gestures and actions produced by the combination of human hands or hands and arms. To recognize and track gestures in a high-dimensional observation space, it is necessary to There is a lot of gesture feature information to be processed. When gesture recognition training is performed in the early stage or gesture recognition is performed later, there are often too many gesture feature information data, resulting in high processing complexity and low processing efficiency.

Images