An image processing method, system, device and medium based on deep learning

Abstract

The invention discloses an image processing method based on deep learning, which includes: acquiring a multi-frame depth image to be processed; using a feature coding module based on a deep convolution network to encode the multi-frame depth image to be processed to obtain corresponding information coding Vector; Utilize the information fusion module based on the RNN cyclic neural network to fuse the information encoding vector to obtain the fusion encoding vector; decode the fusion encoding vector through the feature decoding module based on the depth deconvolution network to obtain the point cloud map, Wherein, each pixel point in the point cloud image is a three-dimensional data point. The present invention uses a deep neural network to replace traditional complex calculations, and can realize real-time reconstruction fusion. In addition, the deep neural network is used to extract features and perform face stitching. After training with hundreds of millions of different expression change data, it can robustly cope with various facial deformations, various angle changes, and various incomplete situations.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to an image processing method, system, device and medium based on deep learning. Background technique [0002] With the rise of the artificial intelligence trend, 2D face recognition technology based on two-dimensional images has developed rapidly and matured, and has become an important means of personal identification. However, two-dimensional face recognition has a high error rate in scenes such as uneven lighting, insufficient lighting, large face angles, and makeup, and is easily attacked by two-dimensional images and videos. Therefore, people began to study 3D face technology to overcome these problems. question. [0003] However, there are still many problems to be solved for the current 3D face recognition technology based on 3D information. Currently limited by the performance of 3D cameras, the quality of single-frame facial depth images collected by 3D cameras is...

AI Technical Summary

Problems solved by technology

However, the existing algorithms generally use the ICP algorithm with a very large amount of calculation, and the incomplete face lacks outstanding geometric features, which makes the ICP algorithm very unstable, and the ICP-based algorithm will gradually accumulate errors. In order to eliminate these problems, it is necessary to introduce More complex optimization algorithms, which result in huge computational overhead, which makes it impossible to achieve real-time multi-frame splicing and fusion in engineering applications, especially on mobile devices

[0005] In addition, when shooting multiple frames, the expression of the face often changes, which cannot be dealt with by traditional algorithms. The change of expression will not only lead to the failure of splicing, but even if it can be spliced, the changing surfaces will be superimposed and become chaotic. invalid data

Images