SSD face detection method based on deep learning

Abstract

The invention discloses an SSD face detection method based on deep learning, and the method comprises the steps: carrying out the candidate box extraction, target detection and bounding box regressionof to-be-detected face data based on a deep convolutional neural network; discretizing an output space of the target Bounding Box regression task into a priori box by utilizing an SSD strategy, and respectively setting priori boxes with various aspect ratios and sizes corresponding to each position in each detection layer; wherein the deep convolutional neural network generates a confidence coefficient of the to-be-measured face according to the priori frame, and corrects the position of the priori frame; and in combination with an SSD strategy, carrying out detection of a plurality of feature maps with different resolutions on the corrected position to obtain a final detection result. According to the SSD strategy provided by the invention, the difference between detection layer data isincreased, the generalization ability and robustness of the model are improved through data enhancement, the problem of difficult negative sample mining is solved, the purpose of balancing training set category distribution is achieved, and the precision and efficiency of small face detection are improved.

Description

technical field [0001] The invention relates to the technical field of face detection, in particular to an SSD face detection method based on deep learning. Background technique [0002] Traditional face detection algorithms such as VJ based on sliding windows use the strategy of cascading AdaBoost for face region screening and face detection. Since this type of model has fewer parameters and features and simple calculation, it can meet the requirements of real-time detection. However, due to the shortcomings of poor stability and few features in the artificially designed features used in the traditional model, the traditional algorithm cannot detect the diversity of faces, and it is difficult to detect blurred and occluded faces. Therefore, the detection accuracy of the traditional algorithm is not ideal. The R-CNN series of algorithms are two-stage general object detection algorithms based on convolutional neural networks. These algorithms use self-learning CNNs for image ...

AI Technical Summary

Problems solved by technology

However, due to the inherent defects of the detection mechanism of the R-CNN series of algorithms, the huge amount of parameters and calculations of the convolutional neural network, the detection speed of the R-CNN series of algorithms is very slow, and even the fastest Faster R-CNN cannot meet the requirements of real-time detection.

Images