Neural network training method and system

Abstract

The embodiment of the invention provides a neural network training method and system, and the method comprises the steps: for the pixel point in the initial reflection laser image, acquiring projectorrays according to the first spatial intersection point and the central point of the speckle projector if it is judged that a first spatial intersection point exists between an infrared camera projection ray corresponding to a pixel point and a target object; based on the first spatial intersection point, the second spatial intersection point, the projector ray, the virtual plane and the referencespeckle pattern, obtaining an intensity value of the pixel point; obtaining a final object speckle pattern according to the intensity values of the pixel points, and obtaining a synthetic data set according to the final object speckle pattern; and training the neural network by using the synthetic data set. According to the embodiment of the invention, under the condition of less training data, the neural network is trained by generating the synthetic data set, so that the training precision of the neural network is improved, and the requirement of large-scale data volume of the deep learningnetwork can be met.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a neural network training method and system. Background technique [0002] With the continuous development of depth perception technology and artificial intelligence technology, the depth imaging method based on monocular structured light, with its strong learning ability and generalization ability, has a strong competitive performance. [0003] For monocular structured light imaging systems, research and practice have proved that the performance of 3D modeling depends on the quality and scale of data, so the 3D target modeling method based on monocular structured light depth imaging systems requires large-scale and high-quality data , each set of data required includes: an object speckle image, a reference speckle image, and ground truth (GroundTruth, GT for short) data of the disparity between the object speckle image and the reference speckle image. At this time, beca...

AI Technical Summary

Problems solved by technology

[0008] Embodiments of the present invention provide a neural network training method and system, which are used to solve the defect that the neural network cannot be trained when the amount of data is small in the prior art, realize the generation of a synthetic data set, and train the neural network based on the synthetic data set. Perform training to improve training accuracy

Images