A semi-supervised optical flow learning method based on dilated convolutional stacked network

Abstract

The invention provides a semi-supervised learning optical flow method based on a convolutional neural network, which belongs to the field of network design. The method provided by the present invention can be trained for mixed data with labels and unlabeled, and an occlusion-aware loss function is designed, which combines the endpoint error cost function for supervised learning with the data item and smoothing item for unsupervised learning In combination, a semi-supervised learning optical flow model is constructed, a stacked network structure is used in the network architecture, hole convolution is introduced in the convolution layer to increase the receptive field, and an occlusion perception layer is designed to estimate the occlusion area. The network can end-to-end Semi-supervised optical flow learning. The method provided by the present invention can improve the accuracy of optical flow estimation, and also proposes an occlusion-aware loss function to semi-supervised training the network, and designs a stacked network structure on the network architecture to further improve network performance.

Description

technical field [0001] The invention provides an optical flow estimation method, specifically relates to a semi-supervised optical flow learning method based on a hole convolution stacking network, and belongs to the field of network design. Background technique [0002] Optical flow estimation can be regarded as a supervised learning problem. The supervised learning method based on convolutional neural network has achieved good results in solving the optical flow estimation problem, but there are still many problems in the supervised learning optical flow method. First of all, it is difficult to obtain the true value of real-world data. The lack of a large amount of labeled data has become the limitation of the supervised learning optical flow method. Secondly, in order to avoid the loss of motion information, many existing fully convolutional network architectures do not perform pooling operations. However, volume The product operation will still lose the detailed informat...

AI Technical Summary

Problems solved by technology

First of all, it is difficult to obtain the true value of real-world data. The lack of a large amount of labeled data has become the limitation of the supervised learning optical flow method. Secondly, in order to avoid the loss of motion information, many existing fully convolutional network architectures do not perform pooling operations. However, volume The product operation still loses the detailed information of the image, and these problems are still fatal for pixel-level tasks.

Images