A system and apparatus for measuring differences between video consecutive frame and their convolution characteristic maps

Abstract

A system and apparatus for measuring differences between video successive frame and their convolution characteristic maps, belongs to the video understanding field in the computer vision application,In order to solve the problem of increasing the kinds of information available to convolution neural networks, so as to increase the understanding ability of convolution neural network for video data,comprises a camera and a computer, wherein the camera is used for shooting video; the computer stores a plurality of instructions, which are suitable for loading and executing by a processor; the video continuous frame data captured by the camera and the corresponding convolution characteristic map are used for calculating to obtain the difference of time information between the video continuousframe data and the convolution characteristic map; The difference of time information is regarded as a part of the loss function of convolution neural network, and it participates in the gradient descent process of convolution neural network back propagation, so that the gradient parameters of convolution kernel are updated toward the case of retaining the time information of input data.

Description

technical field [0001] The invention belongs to the field of video understanding in computer vision applications, in particular to a method, system and device for measuring the difference between continuous video frames and their convolution feature maps. Background technique [0002] Deep learning uses the model constructed by the neural network structure to realize the end-to-end application mode. At the same time, the storage capacity of the model itself for key information in huge data ensures the reliability of the model, which makes the deep learning model unique compared with traditional algorithms. The advantage of comparison has been studied by many scholars in the fields of image, speech and text in a short period of several years and has made great progress. [0003] In computer vision technology such as target detection, target classification, target recognition, target segmentation and other single-frame image applications, deep learning can obtain corresponding...

AI Technical Summary

Problems solved by technology

However, since the development time of the direction of video understanding is not long, the accuracy in practical application scenarios is not satisfactory.

More and more scholars believe that the existing methods cannot accurately extract the time information of continuous frames of video, which leads to the accuracy of the model is not enough to meet the application requirements, and further improvement of the original method is needed

Images