Optical Flow Estimation and Denoising of Video Images Based on Deep Learning Models

Ang Li,Baoyu Zheng,Chen Zhang,Lei Li

doi:10.1109/access.2020.3014345

Abstract

In order to effectively extract image features highly related to visual perception quality, improve the image quality evaluation method, under the framework of deep learning, combining the optical flow method and the edge detection algorithm, a multi-feature fusion motion based on improved optical flow is proposed target detection algorithm. First, a video fluid model is proposed. The fluid model decomposes the video object area changes into sub-area zoom, rotation and translation movements. The rigid body area and the area hierarchy describe the spatial relationship of pixels, and the rigid body motion describes the time domain relationship of pixels. It provides a region-based video processing. The associated spatiotemporal is association method. Secondly, a video fluid model is proposed. The video fluid model treats all pixels of the same surface imaged in the video as a fluid, using streamlines to represent the regional motion of the object, and streamlines to represent the pixel motion of the video object, using rotators and translation lines to simplify the streamlines when necessary. The streamline of the same fluid is smooth in the time domain, and the flow pattern is smooth in both the time domain and the space domain. Finally, the top-down deep learning generation model conversion is carried out, and finally through continuous adjustment between different levels, the generation model can reconstruct the original sample with lower error, so that the essential characteristics of this sample are obtained, namely the highest abstract representation of the depth model. After processing the deep learning model, the sample features after dimensionality reduction can be obtained, and the recognition module is used on this basis. Experiments show that the optical flow estimation method based on deep learning and multi-grid, optical flow field estimation method based on variational model and desiccation method proposed in this paper are effective, and it is suitable for moving image analysis, target tracking and 3D reconstruction Such research has certain theoretical significance and practical application value.

Highlights

Optical flow refers to the speed of motion of the grayscale pattern in the image
In order to further improve the accuracy of the optical flow model and solve the solution of large displacement optical flow, the algorithm-based variational optical flow field estimation method is used on the basis of the deep learning model, the local optical flow algorithm is integrated into the deep learning model, and each anisotropic diffusion and bilateral filtering techniques are used to improve the accuracy of optical flow estimation
The deep learning local optical flow algorithm is integrated into the deep learning model, and using anisotropic diffusion and bilateral filtering technology, the extended model has more robust anti-noise performance on the basis of maintaining the advantages of the deep learning model, and It can effectively solve the problem of large displacement; in the solution process, the structural texture decomposition method and the coarse to fine pyramid method are used to improve the accuracy of optical flow calculation

Summary

INTRODUCTION

Optical flow refers to the speed of motion of the grayscale pattern in the image. Under the illumination of the light source, the grayscale of the surface of the object presents a certain spatial distribution, which is called the grayscale. A. Li et al.: Optical Flow Estimation and Denoising of Video Images Based on Deep Learning Models transportation, meteorological and other fields. The neural dynamics method uses the neural dynamics model (based on the neural network) to simulate the function and structure of the biological vision system to achieve the purpose of optical flow estimation. At present, this technology is not yet mature [15]. Based on the motion-compensated video denoising method separately uses the relationship between the time and space of the video, this use is separate and cannot fully reflect the essence of the spatio-temporal correlation of the video signal. In order to further improve the accuracy of the optical flow model and solve the solution of large displacement optical flow, the algorithm-based variational optical flow field estimation method is used on the basis of the deep learning model, the local optical flow algorithm is integrated into the deep learning model, and each anisotropic diffusion and bilateral filtering techniques are used to improve the accuracy of optical flow estimation

VIDEO IMAGE RECOGNITION ALGORITHM BASED ON DEEP LEARNING

Findings

CONCLUSION