Research On Optical Flow Estimation Network Based On Multi-scale Attention Mechanism

In recent years,the optical flow estimation method based on deep learning has gradually become the mainstream in the field of optical flow estimation.However,there are still some thorny problems in the existing methods,such as the influence of bad weather,motion blur,small object movement,object occlusion,etc.,and it is not easy to solve these problems with deep learning.When we humans observe things,we often have a certain prior knowledge of the whole thing and the surrounding environment.This is a subjective cognition we have formed in our long-term life.This prior cognition can help us infer the occlusion in the image Content,understand the near size and far size of objects on the image.Based on this prior cognition of human beings,this thesis introduces a multi-scale attention mechanism to study the optical flow estimation network from three perspectives:small object detection,different scale object detection and occlusion detection,aiming to use multi-scale attention.The prior knowledge provided by the force mechanism guides the learning of the network.Specifically,the main work of this thesis includes the following three parts:(1)An attention-aware multi-scale optical flow estimation algorithm is proposed.Aiming at the poor optical flow estimation performance of the optical flow estimation network for small objects,this thesis extracts coarse-scale and fine-scale features in parallel,and combines spatial attention and channel attention on the extracted features to generate multi-scale Optical flow estimation results.By using the fine-scale optical flow estimation results to guide the coarse-scale estimation results,the optical flow estimation ability of the network for small objects is enhanced.(2)A scale-adaptive multi-scale optical flow estimation algorithm is proposed.In view of the problem that the existing network cannot grasp the objects of different sizes in the data set well,this thesis proposes to use the feature selectable module to generate a multibranch network.The feature selectable module is embedded with branch attention,so that each layer of the network can automatically adaptively choose the receptive field.By using multi-branch network paths to construct multi-scale correlation volumes,the optical flow estimation performance of the network for objects of different sizes can be improved to a certain extent.(3)An occlusion-aware multi-scale optical flow estimation algorithm is proposed.Aiming at the problem that the optical flow estimation network fails to estimate the optical flow of the occluded area in the image,this thesis proposes to use the self-attention mechanism to capture the global motion features for the first frame of the image,and reduce the selfattention through the asymmetric non-local similarity module.Computational complexity improves the performance of optical flow estimation in occluded areas.