Key point positioning method

Abstract

A key point positioning method, comprising the following steps: inputting a picture, using a pre-trained full convolutional network to obtain a response map of each key point of an object of interest in the picture; based on the response map of each key point, using a pre-trained The point distribution model obtains the initial location of each key point; based on the weighted constrained mean shift method, iteratively adjusts the location of each key point, and finally obtains the final location of each key point. The present invention organically combines data-driven expression ability and model-driven prior reasoning ability. The fully convolutional network it uses can effectively deal with the rigid and non-rigid transformation of objects in the picture, and the point distribution model can effectively deal with the existence of objects in the picture. The weighted constrained mean shift can reasonably balance the effects of the former two, thus greatly improving the robustness of key point positioning.

Description

technical field [0001] The invention belongs to the technical fields of pattern recognition, computer vision, digital image processing, etc., and more specifically relates to a key point positioning method. Background technique [0002] Key point location is one of the important links in computer automatic processing of images. It aims to quickly and accurately locate the key points of objects of interest in the image with strong semantics, such as the corners of the eyes, nose and mouth in face images. [0003] As far as facial key point positioning is concerned, in the prior art, both the traditional model-driven method and the emerging data-driven method can better deal with near-frontal face pictures with little change in expression and slight occlusion. However, limited by the expressive ability of the model-driven method and the sensitivity of the data-driven method to abnormal points, the existing technology cannot handle face pictures with large head poses, exaggerat...

AI Technical Summary

Problems solved by technology

However, limited by the expressive ability of the model-driven method and the sensitivity of the data-driven method to outliers, the existing technology cannot handle face images with large head postures, exaggerated expressions, and severe occlusions.

In addition, the existing technology usually initializes key point positioning based on the output of the face detector. This initialization strategy makes the positioning algorithm easy to fall into a local minimum, which may even cause the algorithm to fail completely in extreme cases.

Images