Face detection method based on cascaded connection convolutional neural network

Abstract

The invention discloses a face detection method based on a cascaded connection convolutional neural network. The method is realized by two steps of training and testing. Firstly, image preprocessing is carried out, an image to be tested is subjected to scale transformation, and the test image is input into a first-hierarchy network. Then, in a first hierarchy, a full convolution network is adoptedto generate face candidate boxes. On a second hierarchy, a non-maximum value inhabitation method is adopted to further filter the obtained face candidate boxes. Finally, in a third-hierarchy network,the face box is screened and further regressed, and the face is filtered for the last time. According to the method, the compact neural network is adopted, image data features are enhanced through acascade connection network method, image noise is lowered, and a good effect is achieved on the aspects of accuracy and speed.

Description

technical field [0001] The invention relates to the field of face detection, in particular to a face detection method based on a cascaded convolutional neural network. Background technique [0002] As a popular computer vision processing method today, convolutional neural network can automatically select the optimal feature extraction operator. It has already shined in the field of target detection. Face detection, as a branch of target detection, has also been improved accordingly. This paper proposes a face detection method based on a cascaded convolutional neural network, which uses a compact neural network. Compared with the traditional convolutional neural network, it achieves good results in accuracy and speed through the cascaded network method. [0003] The first step of the face detection method usually needs to segment the image and nominate candidate windows. The sliding window method is the general choice in the face detection method. This paper abandons the face...

AI Technical Summary

Problems solved by technology

[0004] Usually, only the candidate window categories are divided into two categories: face area and non-face area, and less consideration is given to the area with some faces. In actual operation, when generating data, it is easy to generate a large number of samples containing only the background and There are some face data, but the positive sample data containing all faces is relatively scarce, and it is impossible to make full use of the data with some faces

Images