Depth-auto-encoder-based human eye detection and positioning method

Abstract

The invention discloses a depth-auto-encoder-based human eye detection and positioning method. The method comprises: for all images with calibrated human eye rectangular frame positions in a training set, binary label graphs are generated by using the calibrated human eye rectangular frame positions; small image blocks are selected randomly from the images in the training set and a plurality of auto-encoders are training with supervision in a layered mode to construct a depth auto-encoder, and the depth auto-encoder is initialized based on weights of all layers of the auto-encoder; small original image blocks and small label image blocks are selected randomly at same positions of the original images and the label graphs, wherein the small label image blocks are used as supervision information and the small original image blocks are used as inputs to optimize the depth auto-encoder; and a plurality of small to-be-measured image blocks are generated at a to-be-measured image in a window sliding mode, small to-be-measured label image blocks of all small to-be-measured image blocks are obtained by using the depth auto-encoder and are combined to obtain a to-be-measured label graph of the to-be-measured image, and binaryzation is carried out on the to-be-measured label graph and a position of a human eye is obtained based on coordinate projection or contour searching.

Description

technical field [0001] The invention relates to the field of pattern recognition and machine learning, in particular to image target detection. More specifically, the present invention relates to a method for human eye detection and localization based on a deep autoencoder. Background technique [0002] The explosive growth of biometric technology applications and the huge demand for transplanting biometric algorithms to embedded and mobile platforms make rapid human eye detection and positioning increasingly important. The traditional target detection algorithm is to construct the feature pyramid of the image, and extract the window slidingly on the pyramid, classify the extracted windows, and finally obtain the position of the target through the non-maximization suppression operation. Although this method has achieved good detection results in simple backgrounds and can achieve real-time performance on ordinary computers, this method cannot handle target detection in comp...

AI Technical Summary

Problems solved by technology

Although this method has achieved good detection results in simple backgrounds and can achieve real-time performance on ordinary computers, this method cannot handle target detection in complex backgrounds well, as well as objects with deformation, viewing angle changes and occlusions. problems, and difficult to achieve real-time on embedded and mobile platforms

In addition, the current target detection method based on the deep convolutional neural network has achieved high detection accuracy, and can handle complex backgrounds and problems with deformation and viewing angle changes very well. However, due to its huge amount of calculation, even with the help of parallel computing technology, it is also difficult to meet the real-time requirements

Images