A method for acquiring moving images by analog acquisition and its application

Abstract

The invention discloses a method for simulating acquisition of moving images and an application thereof. The method comprises the following steps: a 3D modeling software is built in an industrial control computer to carry out 3D modeling of a target scene; Installing ISET software to obtain virtual images of the target scene in the three-dimensional modeling; Based on the experimental device, starting the display screen, turning on the spotlight, adjusting the brightness; Running the ISET software to generate and display the virtual picture training set of the target scene; real-timely acquiring of the target scene training set of virtual images displayed on the display screen by an acquiring camera; optimizing the image training set by ELM adaptive fitting; Based on the optimized trainingset, completing the training of depth learning algorithm; Based on the test set, testing the depth learning algorithm after training to verify the trial training set and ELM adaptive fitting optimization algorithm. By comparison, it can be seen that after eliminating the bias of sensing data, the verification error is significantly reduced.

Description

technical field [0001] The present invention relates to the technical field related to data acquisition, in particular to a method and application for analog acquisition of moving images. Background technique [0002] The effectiveness of computer vision algorithms based on deep neural networks is very dependent on the data set. In conventional machine learning or deep learning, it is often assumed that the distribution of the test set and the training set are consistent, the model is trained on the training set, and then tested on the test set. , the test set is the sample of the real working condition data set, which is used to test whether the trained algorithm can work normally and stably in the target real working condition. [0003] If in the actual problem, the quality of the training set is not high, it will be very different from the test set. At this time, there will be a problem of overfitting: the effect of the model on the test set is not ideal. The similaritie...

AI Technical Summary

Problems solved by technology

[0003] If in the actual problem, the quality of the training set is not high, it will be very different from the test set. At this time, there will be a problem of overfitting: the model does not perform well on the test set.

The similarities and differences between the training set and the test set here are mainly reflected in two parts: the first part is the content and quantity of the data set. Although the content of the actual training set collected here is no problem, the number of samples is too small, because the actual collection of data, processing Data, screening images, and manual labeling will be very time-consuming and laborious. A small team of three or five people cannot complete large-scale actual collection work. The content of the virtual network dataset is not targeted and the content quality is not high

The second part is the collection method of the data set, which is reflected in the actual sensing method. Different camera parameters (position, angle, aperture, focal length, collection frame rate, depth of field, etc.) have a great impact on data collection. If the training set and the test set are collected by different sensing methods, there will be a deviation problem between the two data sets. When reflected in the actual working conditions, there will be an over-fitting problem, resulting in a larger verification error.

Images